Inferring Semantic Similarity from Distributional Evidence: 
an Analogy-based Approach to Word Sense Disambiguation* 
Stefano Federici l 
Simonetta Montemagni 1 
Vito Pirrelli ~ 
IPar.O.La sas, Pisa, ITALY 
21stituto di Linguistica Computazionale CNR, Pisa, ITALY 
Abstract 
The paper describes an analogy-based measure 
of word-sense proximity grounded on distribu- 
tional evidence in typical contexts, and illus- 
trates a computational system which makes use 
of this measure for purposes of lexical disam- 
biguation. Experimental results show that word- 
sense analogy based on contexts of use com- 
pares favourably with classical word-sense 
similarity defined in terms of thesaural proxim- 
ity. 
1 Introduction 
Sense disambiguation of a given word occurrence in a 
specific context (hereafter WSD) requires appeal to a 
wide typology of cues, ranging from syntactic subcate- 
gorization to lexico-semantic information and subject 
domain. In this paper we will focus on the use of lexico- 
semantic information, and will try to tackle the related 
problem of measuring the semantic similarity between 
the surrounding context of the word to be disarnbiguated 
and typical patterns of use of that word in a dictionary 
database. In the literature, semantic similarity is usually 
assessed with reference to a hierarchically structured 
thesaurus (e.g. WordNet, \[Miller, 1990\]). The goal of the 
paper is to investigate an alternative way of measuring 
semantic similarity, based on distributional evidence, 
and to show that this evidence can reliably be used to 
disambiguate words in context. To this end, we will 
make use of textual and lexical resources of Italian: 
nonetheless we are convinced that the general point 
made in this paper has a cross-linguistic validity. 
2 Semantic Similarity and WSD 
Most methods proposed in the literature for establishing 
the semantic similarity of words try to map a given word 
* The work reported in tills paper was jointly earned out by the 
authors in the framework of the SPARKLE (Shallow PARsing and 
Knowledge extraction for Language Engineering) project (LE- 
2111). For the specific concerns of the Italian Academy only, S 
Federiei is responsible for sections 3.2, 3.4 and 3.5, S. Monte- 
magni for 2, 3.3 and 4, and V. Pirrelli for 1, 3.1 and 5. 
in context onto the set of known usages of that word in a 
dictionary database: thesaural information is used as a 
yardstick for measuring the semantic proximity between 
known patterns of use and the context to be disambigu- 
ated. Eventually, the sense supported by those patterns 
which are semantically closer to the context in question 
is selected as the most likely one (see, among others, 
\[Dolan, 1994\], \[Resnik, 1995a, 1995b\], \[Agirre and 
Rigau, 1996\], \[Sanfilippo, 1997\]). 
Suppose that one wants to disambiguate the sense of 
accendere in the verb-object pair accendere-televlsione 
'switch on-tv'. The relevant sense of accendere can be 
inferred-on the basis of known examples such as ac- 
cendere 2-radio 'switch on-radio': this inference is 
supported by any seman/i'c hierarchy where both radio 
and television are specified for the same hyperonym, e.g. 
'device', whether immediate or not. 
Thesaural relationships such as hyperonymy and syn- 
onymy, however, do not always capture the dimension 
of similarity relevant to the context in question. Con- 
sider the verb accendere in the context accendere-pipa 
'light-pipe'. The table below contains typical objects of 
two senses of accendere, 'light' (sense 1) and 
'switch on' (sense 2) as they are attested in the Collins 
Italian-English Dictionary \[1985\], together with the 
objects' corresponding hyperonyms according to a 
monolingual Italian dictionary \[Garzanti, 1984\]. 
90 
verb sense 
accendere_l 
accenderel 
accendere_l 
accendere_l 
accendere 2 
accendere_2 
accendere 2 
Table 
object 
stgaretta 
'cigarette' 
candela 
'candle' 
f~ammoeero 
'match' 
cammo 
'fireplace' 
motore 
'engine' 
lampada 
'lamp' 
radto 
'radio' 
1st 
hyper. 
small roll 
lamp ' 
small suck 
hollow 
device 
source of 
illumination 
receiver 
nth nth+l 
hyper, hyper. 
• > artifact object 
• > artifact Iobject 
> artifact \[object 
• > artifact \[object 
> artifact Jobject 
.. > artifact object 
> artifact Iobject 
The word ptpa, which Garzanti describes as a smoking 
tool, does not match any of the immediate hyperonyms 
of the typical objects of accendere 1 and accendere 2 
in Foutt Onbekende schakelo~ie-instructie.. By 
looking further up in the semantic hierarchy, some 
similarities are indeed found, but they are based on too 
general semantic features to be of avail for discriminat- 
ing among senses 1 and 2 of accendere. 
We suggest that, for accendere-pipa to be understood 
in the appropriate sense, namely accendere 1 as in ac- 
cenderel-sigaretta 'light-cigarette', semantTc proximity 
need be computed on different grounds. The relevant 
similarity with links pipes and cigarettes in this specific 
context relates to their both being typically smoked ob- 
jects, a fact which is orthogonal to their general semantic 
class and can be captured on a distributional basis: p:pa 
and sigaretta are distributionally equivalent relative to 
the same verb sense, i.e. they both occur as typical ob- 
jects of the verb fumare 'smoke'. Distributional equiva- 
lence correlates with semantic similarity under the as- 
sumption that nouns which bear the same syntactic rela- 
tion to the same verb sense are part-of a semantically 
coherent class. It turns out that, in examples such as 
accendere-pipa, distributionally-based semantic simi- 
larities can permit more appropriate sense assignments 
which are specifically tailored to the context to be dis- 
ambiguated. Observe further that also similarities com- 
monly captured on the basis of thesaural information, as 
in the case of radio and televtsione above, can in princi- 
ple be inferred from distributional evidence through 
relevant contexts of use (e.g. spegnere-radlo 
'switch off-radio' and spegnere-televtsione 'switch off- 
tv' in t~ example at hand). 
Summing up, we contend that thesaural relationships 
capture only some of the various dimensions of word sense 
analogy which appear to play a relevant role in the disambi- 
ugation of word co-occurrence patterns. In fact, while the- 
saural relationships are def'med out of context once and for 
all, effective analogies are to be tailored to the specific con- 
textual pattern to be disambiguated. We showed how this can 
be attained on the basis of distributional evidence. 
3 SENSE: a distributionally-based 
WSD system 
SENSE (Self-Expanding linguistic kNowledge-base for 
Sense Elicitation) is a specialised version of a general 
purpose language-learning system (\[Federici and Pirrelli, 
1994\]; \[Federici et el., 1996a\]; \[Montemagni et al., 
1996\]) for carrying out WSD on the basis of distribu- 
tional evidence. 
SENSE's inferential routine requires: 
i) a structured data set of known word co-occurrence 
patterns (WCPs) constituting an Example Base (EB); 
ii) a target context to be disambiguated (TC); 
iii)a best-analogue(s) function (BAF) projecting TC 
onto EB for the best analogue(s) to be selected and 
thus the most likely senses to be identified• 
3.1 Internal architecture of EB 
Word co-occurrence patterns 
WCPs are modelled here as consisting of an input and an 
output level of representation. At the input level, each 
element of the pattern is described by a set of features 
which are expected to be of some use for WSD: lemma, 
part of speech and morpho-syntactic properties (such as 
the syntactic function of nouns with respect to the verb)• 
The output representation simply consists in the ex- 
pected answer, i.e. the sense of each element of the pat- 
tern in the described context. An example of this type of 
linguistic object, illustrating the pattern fumare 1- 
s:garetta_l/O 'smoke-sigarette', is given in Four! (On- 
bekende schakeloptie-instructie.: 
Table 
fumare_l-slgaretta_ llO 
input fumare stgaretta 
verb noun 
object 
output fumare_l sigaretULl 
The input representation is a list of sets of atomic units; 
each feature set (which is assigned a single column in 
the table) describes a distinct element of the pattern. In 
output, the list of atomic units "fumare 1" and 
"sigaretta 1" indicates the senses of the elements in the 
specific context• Elements in the input and output lists 
are conventionally ordered. 
In the current version of Italian EB used for our pur- 
poses, WCPs are verb-noun pairs where the relation of 
the noun to the verb is either subject or object. This 
presupposes a preliminary stage of morpho-syntactic 
parsing \[Montemagni, 1995\]: co-occurrence patterns 
abstract away from actual word forms and are aug- 
mented with information about grammatical relations. 
Note that although availability of pre-processed input 
makes word sense disambiguation simpler and more 
accurate, it is in no way a necessary precondition for the 
task to be carried out. 
91 
Pairwise analogies 
The Italian EB consists of WCPs of the type illustrated 
in Fout! Onbekende schakeloptie-instruetie, above. 
Note however that they are not used as such; rather they 
form part of a distributed network ~ which is constructed 
so as to i) factor out the optimal set of analogies shared 
by all WCPs in EB, and ii) link the found analogies with 
their corresponding complements relative to the full 
WCPs (so-called differing parts). To make this picture 
more concrete, let us consider some simple examples. 
Given a pair of word co-occurrence patterns wcp, and 
wcp2, they are judged to be analogous if they share some 
representation units at both input and output levels. Any 
shared collection of units of both levels is referred to as 
an analogical core (or simply core, written wcplnwcp2 ). 
Suppose that wcp, and wcp2 are fumare 1-slgaretta_l/O 
and fumare_l-pzpa_l/O 'smoke-pipe' respectively, de- 
fined as in Fout! Onbekende sehakeloptie-instruetie. 
above and Fout\[ Onbekende sehakeloptie-instruetie. 
below. 
Table 
input 
output 
fumare l-pipa_l/O 
furnare , ' ptpa 
verb noun 
obJeCt 
fumare_l ptpa_l 
Their core is identified by a function (MF) mapping one 
set of units in Fout! Onbekende sehakeloptie- 
instruetie, onto one set of units in Fout! Onbekende 
sehakeloptie-instruetie, through the identity relation. 
MF is order-sensitive, so that only sets which take the 
same relative order in the lists are mapped onto each 
other. Fout! Onbekende sehakeloptie-instruetie, gives 
a possible result of this operation in the leftmost box 
headed by wcptnwcp2. The core in question is a verb- 
noun pair where the noun element is specified only at 
the input level, for a subset of the features describing the 
noun elements in the compared patterns, while nothing 
being said as to the possible sense interpretation of the 
noun. Nonetheless, the information about the noun con- 
veyed by the core, namely its syntactic relation to the 
verb, is part of the knowledge supporting the interpreta- 
tion of the verb asfumare_l: i.e. the verb in this reading 
is used transitively. 
Table 
input 
wcp~n wcp2 wcprwcp2 ..1 wcp2- wcpl 
fumare sigaretta I plpa 
verb noun \[ 
{object 
1 In the current version of SENSE a (partial) network structure is 
built from scratch every time a new TC is presented to the system 
However, for the sake of clarity, in what follows we Illustrate the 
working of our system as though the network structure were built 
during the acquisition of EB. See \[Federiei et al, 1996b, p.393\] 
for a discussion of the two alternatives. 
output l fumare I I l lsig retta_ll.l l 
The complements of the core relative to wcp~ and wcp2 
designate those units which are specific to the compared 
objects: they constitute the so-called differing parts, 
illustrated in Four! Onbekeude sehakeloptie-instruetie. 
in the columns headed by wcp,-wcp2 and wcp2-wcp, 
respectively. They contain information about the lexical 
fillers of the noun slots of the patterns. 
Network structure of EB 
Cores and remaining parts are always anchored to a 
given pair of linguistic objects: in fact, cores cannot be 
extracted either from existing cores or from existing 
differing pans. When more than one pair of WPCs is 
considered, it may turn out that what is a core relative to 
a given pair is a remaining part relative to another pair. 
Suppose that MF maps fumare_l-stgaretta_l/O (wept) 
onto accendere_l-szgarettal/O (wep3). One of the pos- 
sible results of this mapping is shown in Fout! Onbek- 
ende sehakeloptie-instructie, below: 
Table 
wep3-wep, 
input accendere 
verb noun 
object 
output accendere 1 
wep,-wcps I wcp,~wep3 .... fumare stgaretta 
noun verb ';6jee( 
fum~e 1 "'" slgaretta 1 
Comparison of cores and remaining parts in Fout! On- 
bekende schakeloptie-instructie, and Font! Onbek- 
ende sehakeloptie-instruetie, above shows that one of 
the remaining parts relative to wcp~ and wcp2 (namely 
wcpl-wcp2) is identical to the core relative to wcp, and 
wcp3 (wcplnwcp3). 
The informational content of Fout! Onbekende 
sehakeloptie-instructie, and Foat! Onbekende 
sehakeloptie-instructie, can be represented conven- 
iently through the graph in Fout{ Onbekende 
schakeloptie-instruetie.. 
plpa_l 
fumare_l-object o .~* wcp2-wePl 
wcp~-wcp3 = wcplnwcp2 stgaretta_l 
.~t- wepl-wep2 = wcplnwcp3 accendere l-object 
wcp3-wcpl 
Figure An analogical family 
The graph represents cores and remaining parts as con- 
nected nodes, each accompanied by a mnemonic label. 
For example s:garetta_l corresponds to wcp,-wcp2 = 
wcp,n wcp3. An (unoriented) arc connecting two nodes 
expresses their "complementarity", i.e. the intuitive 
notion that the two connected nodes, taken together, 
cover an attested WCP in its entirety. For instance, fu- 
mare 1-0bject is connected with sigaretta 1 since they 
form together an attested WCP, names fumare 1- 
stgaretta_I/O. By contrast, no direct connection is ~b- 
92 
served between accendere_l-object and plpa_l, to sig- 
nify that no corresponding pattern is attested in EB. 
Remaining parts which are connected with the same core 
are said to be contrastive, since, by replacing one with 
the other, different WCPs are obtained. A graph like the 
one in Foutt Onbekende sehakeloptie-instructie, rep- 
resents in our terms an analogical family (AF). Clearly, 
far more extended AFs than the one in Fout! Onbek- 
ende sehakeloptie-instructie, can be found. 
Among the WPCs of the AF in Fout! Onbekende 
schakeloptie-instructie., fumare_l-sigaretta_l/O is the 
only one which is made up out of two cores, namely 
wcpln wcp3 and wcpln wcp3. Due to its pivotal position 
in the graph, it is some times useful to refer to it as the 
"hook pattern", or more simply "hook", of the AF in 
question. Accordingly, we will call the noun collocate of 
a hook, i.e. stgaretta_l in the example at hand, "hook 
noun", and the corresponding verb, i.e. fumare_l, "hook 
verb". Note further that the hook noun stgaretta_l is 
functional to establishing a kinship between the verb 
senses fumare_l and accendere_l, since it denotes a 
non-empty intersection between typical patterns of their 
use. 
3.2 The Best-analogue(s) Function 
Unlike linguistic objects in EB, which are specified for 
two representation levels (input and output), a Target 
Context (TC) is specified at the input level only, since 
its sense is precisely what the system has to predict on 
the basis of the available knowledge. 
This prediction is carried out through operation of the 
best-analogue(s) function (BAF) which projects TC onto 
EB, searching for TC's best candidate analogue(s). BAF 
uses the notion of distributionally-driven word-sense 
analogy developed in the previous pages, and can be 
informally described through the following steps: 
a) if EB contains a pattern wcp, which fully matches TC 
. at the input level, then wcp, is the best analogue and 
its output is ranked first in the list of available an- 
swers; note that this step does not stop SENSE from 
continuing its search; 
b) if EB contains a single AF such that two of AF's 
nodes together cover TC's input representation in its 
entirety, the output representations associated with 
the matching nodes is added to the list of available 
answers with a ranking score, gauged as a function of 
type and quantity of supporting evidence (see below 
for more detail); 
c) if steps a) and b) yield no result, no output is pro- 
vided by SENSE. 
BAF at work 
Let us look at some interesting cases of BAF at work. 
Note that all examples considered in this paper are rep- 
resentative of real test suites of SENSE, and the assumed 
knowledge in EB reflects the current status of an actual 
data base acquired from typical examples of use within 
verb entries of the Collins Italian-English dictionary 
\[1985\]. 
Suppose that SENSE has to assign a verb sense in the 
target context accendere?-ptpa_l/O 'light-pipe'. The 
context being not attested in EB, TC is projected onto 
EB's network, for a relevant AF to be found. The AF in 
Fout! Onbekende sehakeloptie-instructie, above is a 
good instance of such a relevant family, since it contains 
two nodes, namely accendere_l-object and plpa_l, 
which fully cover TC's input. Step a) having failed, the 
two nodes in question are not directly connected; none- 
theless, their belonging to the same family means that 
there exists a continuous path of complementarity arcs 
joining the two. This continuity allows SENSE to hy- 
pothesize an arc directly connecting accendere_l-object 
with pzpa_l (represented as a dashed line in Font! On- 
bekende schakeloptie-instructie, below): 
fumare_l-objeet o~./"'/e pzpa l 
accendere l-objeet i'/'/~ s:garetta_l 
Figure A reconstructed connection 
i.e. the co-occurrence pattern accendere_l-ptpa_l/O can 
be reconstructed on the basis of the available distribu- 
tional evidence, and supports the interpretation ac- 
cendere 1. 
To sum up, SENSE identifies a distributional similarity 
between accenderel and fumare_l: this similarity is based 
on the fact that cigarettes can both be lit and smoked. This 
triggers the analogy-based inference that pipes, besides being 
smoked, can also be lit, thus supporting the interpretation of 
accendere 1 
3.3 Constraints on distributionaily-based 
WSD 
In the example illustrated above, nouns stand in the 
same syntactic relation to the verbs. However, it is often 
the case that clusters of nouns which function as the 
object of a given verb can function as typical subjects of 
other, somehow related, verbs. If this sort of systematic 
subject/object alternation is taken into account, the in- 
ferential power of distributionally-based WSD may in- 
crease considerably, as shown by the following exam- 
pies. 
Consider the TC attaccare_?-fotografial/O 
'hang_up-photograph'. EB contains three different 
senses of attaccare, each attested with the following sets 
of noun collocates: 
• attaccare_l-\[francobollo_l/O, manifesto_l/O, 
quadro_l /O \] 
' hang_up'- \['stamp/O', 'poster/O', 'painting/O'\] 
• attaccare_2-\[dtscorso_l/O\] 
'start'-\['conversation/O'\] 
• attaccare_4-\[moda_l/S\] 
' catch_on'- \[' fashion/S' \] 
No one of the noun collocates listed above happens to be 
attested in EB as an object of verbs which also combine 
93 
with fotografia as an object. However, if the restriction 
that relevant nouns must stand in the same relation to the 
predicate is relaxed, then relevant distributional evi- 
dence can in fact be found in EB. Fotografia and quadro 
'painting' are both attested as typical subjects of the 
verb rappresentare_l, a fact which can be interpreted in 
terms of Pustejovsky's telic role \[Pustejovsky 1995\], 
since both nouns are normally used to "show some- 
thing". Furthermore, quadro is also attested as a typical 
object of the verb attaccare_l; on this basis, it can rea- 
sonably be supposed that also fotografia, when co- 
occurring as an object of attaccare, points to the sense 
attaccare 1. 
Inferences based on AFs involving asymmetric syn- 
tactic dependencies permit to exploit the data contained 
in EB to the full. Moreover, the procedure becomes es- 
sential for generalising over cases of so-called valency 
alternation. Consider the causative-inchoative alterna- 
tion, which involves two argument structures, a transi- 
tive and an intransitive one: a verb such as aumentare 
'increase' can be used in a sentence like la Fiat ha 
aumentato gh stipendl agh operat 'Fiat increased sala- 
ries to workers', where stipendio is the object of the 
verb, and in a sentence like gh snpendt aumentarono 
inaspettatamente 'salaries increased unexpectedly', 
where sttpendio is the subject. In the literature, the theo- 
retical issue of whether alternating argument structures 
of the same verb should be associated with a unique 
sense or with different senses of that verb is still open. 
In practice, lexicographers' approaches vary considera- 
bly, depending on factors such as the dictionary's inter- 
nal structure or main practical purpose: for instance, in 
bilingual dictionaries different but alternating argument 
structures often give rise to different senses, due to dif- 
ferences in their translation. Whatever approach is 
adopted by the lexicographer, however, SENSE is capa- 
ble of identifying a sense alternation induced by an al- 
ternation of argument structure, or~ alternatively, of 
recognising two different argument structures as related 
to the same verb sense, thanks to its ability to deal with 
asymmetric syntactic dependencies in EB. 
To sum up, word sense disambiguation with verb- 
noun pairs involving asymmetric dependencies is more 
effective than when only contexts with symmetric de- 
pendencies are considered. This procedure is particularly 
crucial for verbs alternating between different arguments 
structures. 
3.4 Beyond attested evidence 
SENSE's inferential routine can go beyond attested evi- 
dence; in fact, the presence of an attested pattern which 
matches exactly TC's input does not prevent the system 
from exploring other hypotheses. This flexibility is often 
useful: when sense distinctions are fine grained and data 
in EB are sparse, distributional criteria get too coarse 
grained to be able to point to a unique sense interpreta- 
tion. 
Consider, for example, battere ?-mano_l/O 'hit- 
hand': in EB, this pattern is attested with the sense of 
clapping, as an instance of beating body parts with a 
regular rhythm (battere_3). However, there is at least 
another sense of battere which is appropriate in the 
context considered, namely battere_l, understood under 
the more general sense of hitting someone or something. 
In cases like this one, SENSE "ambiguates" the verb- 
noun pair received in input, by finding out other plausi- 
ble sense assignments besides the one attested in EB. As 
a consequence, SENSE outputs more than one sense 
interpretation, while ranking the attested interpretation 
first. Identification of alternative sense assignments, 
although with lower ranking, comes in handy when the 
expected TC interpretation is not the attested one. This is 
reasonable, we believe, since WSD is often a matter of 
suggesting a set of more or less plausible interpretations 
in context rather than asserting one interpretation only; 
by taking attested evidence (no matter how representa- 
tive) at face value one would wrongly ignore the com- 
mon fact that, even in real usages, a target context can in 
fact be understood in more than one way. 
3.5 Ranking multiple disambiguation re- 
sults 
As just shown, distributlonally-based word sense disam- 
biguation does not always make the system converge on 
a unique interpretation. This situation typically occurs 
when different senses of a word are close in meaning, 
and this closeness is reflected by their co-occurrence 
with distributionally similar if not identical words. When 
more than one sense interpretation appears to be plausi- 
ble, different strategies can be followed in order to rank 
them from more to less likely. When the set of plausible 
interpretations includes a directly attested one, then the 
latter is always ranked first. Ranking of inferred inter- 
pretations needs to take into account a number of differ- 
ent factors. 
As a first approximation, different sense interpretations can 
be ranked according to the number of AFs supporting them. 
Suppose that SENSE has to assign a verb sense in accarez- 
zare_?-speranza_I/O 'toy_with-hope'. Both possible 
sense interpretations of accarezzare (i.e. accarezzare 1 
'stroke' and accarezzare2 'toy_with') are supported. In EB, 
the interpretation accarezzare_l is supported by one AF only, 
which includes the pattern perderel-capellol/O 'lose-hair'. 
On the other hand accarezzare2 is supported by four AFs, 
each containing the following hooks: 
1. abbandonare_4-progetto_l/O 'giveup-project' 
2. cullare 1-tdea 1/0 'cherish-idea' 
3. nascere 2-tdea_l/S 'be_born-idea' 
4. nau~agare 1-progetto_I/S 'fall~hrough-project' 
Hence, accarezzare2 gets score 4 and wins out over ac- 
carezzare I which scores 1. 
The sheer number of supporting AFs, however, is too gross 
a criterion when used on its own. Consider the target af- 
fluire?-acqua_l/S 'flow-water'. Here, the contextually more 
appropriate sense affluwel 'flow' is supported by three AFs, 
while affluwe2 'pourin' is pointed to by five different AFs: 
af~uire l 
94 
1. intorbtdare_l-liqutda-1/O ' cloud-liquid' 
2. penetrare 2-1iquido_l/S 'percolate-liquid' 
3. versare_2-liqutdo_l/O'pour-liquid' 
affluire_2 
1. conflutre_l-persona_l/S 'join-person' 
2. gettare 1-persona_1/O 'rush_in-person' 
3. imbarcare_l-mercel/O 'ship-goods' 
4. insinuarsi_3-persona_l/S 'creep_into-person' 
5. ristagnare_l-persona_l/S 'lag_person' 
Nonetheless, SENSE could be "persuaded" to prefer the cor- 
rect interpretation if also the typology of supporting evidence 
is taken into account. Intuitively, preference has to be given to 
more specific supporting semantic evidence over semantically 
vaguer one. In our terms, this means that supporting AFs 
which contain a more specific hook noun should carry more 
weight for WSD than AFs containing vaguer hook nouns. 
Usually, generality of a word is measured by referring to 
a semantic hierarchy. In this context we have used fre- 
quency of word occurrence in EB as a convenient measure of 
"generality/specificity" of a word: the more often a hook 
noun occurs as a subject/object of differents verbs in EB, the 
more general it can be considered. Note that EB contains only 
WCP types, so that word counting here is significantly differ- 
ent from counting token frequencies in a real text; type fre- 
quency appears to point more decisively to the general struc- 
ture of lexical competence, rather than to distributional effects 
in language performance. On this basis, each relevant AF is 
assigned a specificity score, equal to the inverse ratio of the 
number of times its hook noun occurs in EB. The ranking 
score of a given sense interpretation S is then the sum of the 
specificity scores of all AFs = { AFt, AF~ ..... AF, } support- 
ing it: 
Ss~>~,~ = Spec(AF0+ Spec(AF2)+...+ Spec(AF,) 
where Spec(AF,)= 1/type-frequency(hook_noun). 
ha the light of this score, :ranking of the senses of affluwe is 
reversed: the best disambiguation hypothesis is now aJ: 
flutre_l (ranking score 0.281046), against afflutre_2 whose 
ranking score 0.069598 is significantly lower. The hook noun 
supporting the sense affluire_l is hqutdo 'liquid', whose 
specificity score is 0.111111 when used as an object and 
0.058824 when used as a subject. By contrast, the same score 
is significantly lower in the cases supporting the other sense: 
0.007194 for persona_l/O and 0.005650 for personal~S; 
0.045455 for merce 1/0. 
The specificity score tends to overrate very specific analo- 
gies, that is analogies supported by analogical families with 
highly idiosynractic collocates, over more general analogies. 
To counterbalance this bias, another ranking factor, called the 
"coverage" score, can usefully be exploited in our context. 
For each available sense interpretation of TC attested in EB, 
we count how many of its collocates occur as hook nouns of 
all AFs supporting that sense. Note that, for an AF to support 
a certain verb sense, it has to contain as a hook noun a collo- 
cate of the verb sense S in question. We then assign to S a 
coverage score S~ov~:~o~, which is proportional to the num- 
ber of shared collocates: 
S,~w~e ~ = # nour~collocate(AF(S))/#noun collocate(S) 
where '# nout~collocate(AF(S))' reads "cardinality of the 
noun collocates of the AFs supporting S", and 
'# nouncollocate(S)' reads "cardinality of the noun collo- 
cates of S". The bigger this score, the more widely supported 
the corresponding sense interpretation in EB. This follows 
quite naturally in an analogy-based perspective, since, intui- 
tively, two verb senses are considered more similar if they 
have more collocates in common. Eventually, this score is 
combined with the other scores considered above to yield a 
fmal ranking score: 
S lo~1 ~nS_~or~ = Ssr~:o~ x S~ov~gn_~r~ 
To give a concrete example, assume that SENSE has to inter- 
pret the pattern accostare?-qualcuno_l/O 'approach- 
somebody'. Accostare is attested in EB in three different 
senses: accostare_l 'bring_near' with words like chair, object 
and ladder, among its typical objects; accostare2 'approach' 
with person as typical object; accostare3 'setajar' said of 
shutter and door. If the coverage score is not considered, the 
ranking would be accostare 3 (0.428571), accostare 1 
(0.316417), accostare2 (0.12~02) the latter being the ap- 
propriate sense in this context. Intuitively this is due to the 
fact, that, for example, the AFs supporting accostare 3 all 
exhibit one hook noun only, namely porta, which none~eless 
contributes a high specificity score, due to its poor type- 
frequency in EB. Yet, if the coverage score is taken into ac- 
count, the ranking becomes accostare 2 (2.079136), ac- 
costarel (1.582087), accostare3 (0.4~571), with the ap- 
propriate sense ranked first. 
4 SENSE: experimental results 
Experiments have been carried out with an EB of 8,153 
distinct verb-noun patterns (2,488 verb-subject, 5,665 
verb-object) automatically extracted from the whole set 
of verb entries of the Collins bilingual Italian-English 
dictionary \[Montemagni, 1995\]. In these patterns only 
verbs are disambiguated as to their sense, whereas nouns 
are assigned all possible senses. These patterns exem- 
plify 3,359 different verb senses, each illustrated, on 
average, through 2.42 patterns. In Font! Onbekende 
schakeloptie-instructie, below, verb senses are ranked 
per number of exemplifying patterns: 
Table 
n of patterns verb senses 
10-15 21 0 6% 
9-6 188 5 6% 
2-5 1874 55 8% 
1 t276 38% 
total 3359 100% 
Senses which are attested in ten or more patterns are a 
negligible part of EB; actually, most verbs are illustrated 
by means of a number of patterns ranging between 2 and 
5. Finally, a considerable group of verb senses is attested 
only once. Note that this does not stop SENSE from 
recognising them in unseen contexts; e.g. in EB there is 
95 
only one pattern exemplifying the verb sense abbas- 
sare_3 'reduce' (namely, abbassare_3-prezzo/O), but 
this does not prevent SENSE from recognising it in tar- 
get contexts such as abbassare_?-sttpendzo/O. 
SENSE's performance has been tested on a corpus of 
150 TCs randomly extracted from unrestricted texts. 
Patterns which already occur in EB were excluded from 
the test corpus since we wanted to focus on the reliabil- 
ity of inferences based on distributional evidence, rather 
than on EB's statistical representativity. The results of 
this experiment are reported below: 
Table 
Polysemous 
~I~ECISION 
Overall 
~ECALL i 793% 66 3% 
i 89 9% 80 4% 
Figures in the first column refer to both polysemous and 
monosemic verbs; here, recall and precision are high and 
refer to the topmost sense in the ranking only. In the 
second column, recall and precision are relative to 
polysemous verbs only, and in spite of an obvious de- 
crease compare well with related work carried out with 
different methods (see, for instance, \[Agirre and Rigau, 
1996\]), and are in fact very promising if one considers 
the comparatively small size of EB, and that only part of 
its attested words are semantically disambiguated. 
5 Concluding remarks 
In this paper we described a WSD system which uses a 
notion of semantic similarity based on distributional 
evidence. Prehminary results look promising. 
The described measure of semantic similarity offers 
significant advantages compared with methods where 
word similarity is evaluated either in statistical terms, 
ultimately based on token frequency, or through refer- 
ence to a hierarchically structured thesaurus. First, good 
results are achieved with small quantities of data, part of 
which are not even semantically disambiguated. Second, 
the suggested measure is sensitive to similarities'which 
are relevant to the context being disambiguated, thus 
overcoming one of the major drawbacks of fixed decon- 
textualised semantic hierarchies. 
On a more practical front, this measure was evaluated 
as an integral part of the disambiguation strategy of 
SENSE, whose main advantages over other WSD sys- 
tems can be summarised as follows: 
• SENSE does not take attested evidence at face value 
but always entertains other hypotheses; 
• SENSE's inferences are not restricted to contexts 
which exhibit symmetric syntactic dependencies, but 
also exploit alternations in argument surface realisa- 
tion with semantically related verbs; 
• SENSE is sensitive to the semantic general- 
ity/specificity of supporting evidence. 

References 
\[Agirre and Rigau, 1996\] E. Agirre, G. Rigau. Word 
Sense Disambiguation using Conceptual Density. Pro- 
ceedmgs of COL1NG-96, Copenhagen, Denmark, pp. 16- 
-22, 1996. 
\[Collins, 1985\] Collins Giunti Marzocco. English-Italian 
Itahan-English Dictwnary. Collins Giunti Marzocco, 
London Firenze, 1985. 
\[Dolan, 1994\] W. Dolan. Word Sense Ambiguation: 
Clustering related Senses. Proceedings of COLING-94, 
Kyoto, Japan, pp. 712--716, 1994. 
\[Federici and Pirrelli, 1994\] S. Federici, V. Pirrelli. Lin- 
guistic Analogy as a Computable Process. Proceedings 
of NeMLaP, Manchester, UK, pp. 8--14, 1994. 
\[Federici et al., 1996a\] S. Federici, S. Montemagni, V. 
Pirrelli. Analogy and Relevance: Paradigmatic Networks 
as Filtering Devices. Proceedmgs of NeMLaP, Ankara, 
Turkey, pp. 13--24, 1996. 
\[Federici et al., 1996b\] S. Federici, V. Pirrelli, F. Yvon. 
A dynamic Approach to Pardigm-driven Analogy. in S. 
Wermter, E. Riloff, G. Scheler (Eds.) Connectwmst, 
Stattstwal, and Symbohc Approaches to Learnmg for 
Natural Language Processing, pp.385-398, Springer, 
1996. 
\[Garzanti, 1984\] Garzanti. I1 Nuovo Dtztonarto Itahano 
Garzanti. Garzanti, Milano, 1984. 
\[Miller, 1990\] G. Miller. Ftve Papers on WordNet. Spe- 
cial Issue of International Journal of Lexzcography, 
3(4), 1990. 
\[Montemagni, 1995\] S. Montemagni. SubJect and ObJect 
m Itahan Sentence Processing, PhD Dissertation, 
UMIST, Manchester, UK, 1995. 
\[Montemagni et al., 1996\] S. Montemagni, S. Federici, V. 
Pirrelli. Resolving syntactic ambiguities with lexico-semantic 
pattems: an analogy-based approach Proceedings of 
COLING-96, Copenhagen, August 1996, pp. 376--381, 1996. 
\[Pustejovsky, 1995\] J. Pustejovsky. The Generative 
Lexwon. The MIT Press, Cambridge, Massachusetts, 
1995. 
\[Resnik, 1995a\] P. Resnik. Using Information Content to 
Evaluate Semantic Similarity in a Taxonomy. Proceed- 
rags of lJCA1-95, 1995. 
\[Resnik, 1995b\] P. Resnik. Disambiguating noun 
groupings with respect to WordNet senses. Proceedings 
of 3rd Workshop on very large corpora, Association for 
Computational Linguistics, 1995. 
\[Sanfilippo, 1997\] A. Sanfilippo. Using Semantw Simt- 
larzty to Acquire Cooccurrence Restrwtions from Cor- 
pora, SPARKLE Project (LE 2111), Working paper 
n,12, 1997. 
